On Wed, May 20, 2015 at 10:37:52AM -0700, Scott Feldman wrote:
> On Wed, May 20, 2015 at 5:57 AM, Simon Horman
> <simon.hor...@netronome.com> wrote:
> > On Wed, May 20, 2015 at 01:17:36PM +0200, Jiri Pirko wrote:
> >> Wed, May 20, 2015 at 10:46:26AM CEST, simon.hor...@netronome.com wrote:
> >> >On Wed, May 20, 2015 at 05:36:06PM +0900, Toshiaki Makita wrote:
> >> >> On 2015/05/20 16:48, Simon Horman wrote:
> >> >> > On Wed, May 20, 2015 at 03:15:23PM +0900, Toshiaki Makita wrote:
> >> >> >> On 2015/05/20 14:48, Simon Horman wrote:
> >> >> ...
> >> >> >>> static void _rocker_neigh_add(struct rocker *rocker,
> >> >> >>> + enum switchdev_trans trans,
> >> >> >>> struct rocker_neigh_tbl_entry *entry)
> >> >> >>> {
> >> >> >>> + if (trans == SWITCHDEV_TRANS_PREPARE)
> >> >> >>> + return;
> >> >> >>> entry->index = rocker->neigh_tbl_next_index++;
> >> >> >>
> >> >> >> Isn't index needed here? It looks to be used in later function call
> >> >> >> and
> >> >> >> logging.
> >> >> >
> >> >> > Thanks, that does not follow the usual model of setting values
> >> >> > during the PREPARE (and all other) transaction phase(s).
> >> >> >
> >> >> >> How about setting index like this?
> >> >> >>
> >> >> >> entry->index = rocker->neigh_tbl_next_index;
> >> >> >> if (trans == PREPARE)
> >> >> >> return;
> >> >> >> rocker->neigh_tbl_next_index++;
> >> >> >> ...
> >> >> >
> >> >> > I am concerned that _rocker_neigh_add() may be called by some other
> >> >> > caller while a transaction is in process and thus entry->index will
> >> >> > be inconsistent across callers.
> >> >> >
> >> >> > Perhaps we can convince ourselves that all the bases are covered.
> >> >> > So far my testing has drawn a blank. But the logic seems difficult to
> >> >> > reason about.
> >> >> >
> >> >> > As we are basically allocating an index I suppose what is really
> >> >> > needed for
> >> >> > a correct implementation is a transaction aware index allocator, like
> >> >> > we
> >> >> > have for memory (rocker_port_kzalloc etc...). But that does seem like
> >> >> > overkill.
> >> >> >
> >> >> > I think that we can make entry->index consistent across
> >> >> > calls in the same transaction at the expense of breaking the
> >> >> > rule that per-transaction data should be set during all transaction
> >> >> > phases.
> >> >> >
> >> >> > Something like this:
> >> >> >
> >> >> >
> >> >> > if (trans != SWITCHDEV_TRANS_COMMIT)
> >> >> > /* Avoid index being set to different values across calls
> >> >> > * to this function by the same caller within the same
> >> >> > * transaction.
> >> >> > */
> >> >> > entry->index = rocker->neigh_tbl_next_index++;
> >> >> > if (trans == SWITCHDEV_TRANS_PREPARE)
> >> >> > return;
> >> >> >
> >> >> >
> >> >>
> >> >> As long as it is guraded by rtnl lock, no worries about this race? It
> >> >> seems to be assumed that prepare-commit is guarded by rtnl lock,
> >> >> according to commit c4f20321 ("rocker: support prepare-commit
> >> >> transaction model").
> >> >>
> >> >> But as you are concerned, it seems to be able to be called by another
> >> >> caller, specifically, neigh_timer_handler() in interrupt context without
> >> >> rtnl lock. IMHO, it should be fixed rather than avoiding the race here.
> >> >
> >> >Yes, I believe that is the case I was seeing.
> >> >
> >> >Scott, Jiri, how would you like to resolve this?
> >>
> >>
> >> I believe that you can depend on rtnl being held - in
> >> switchdev_port_obj_add
> >> there is ASSERT_RTNL assection at the very beginning of the function.
> >
> > In the prepare-commit scenario, yes, I agree that is the case.
> > But it does not seem to always be the case when the transaction phase is
> > none.
> >
> > What I am seeing is:
> >
> > 1. rocker_port_ipv4_nh() is called via switchdev_port_obj_add()
> > with trans = SWITCHDEV_TRANS_PREPARE
> >
> > 2. rocker_port_ipv4_neigh() is called by rocker_neigh_update()
> > with trans = SWITCHDEV_TRANS_NONE.
> >
> > The call chain goes up to arp_process() via neigh_update().
> >
> > 3. rocker_port_ipv4_nh() is called via switchdev_port_obj_add()
> > with trans = SWITCHDEV_TRANS_COMMIT
> >
> > I believe #2 is not guarded by rtnl.
>
> Looks like rocker->neigh_tbl_next_index was a problem even before the
> transaction model was introduced, due to no protection for concurrent
> processes in diff contexts.
>
> We'll need to turn the NETEVENT_NEIGH_UPDATE into process context and
> hold rtnl_lock, similar to what we do in
> rocker_event_mac_vlan_seen_work(). That, plus Toshiaki's suggested
> change for _rocker_neigh_add() should do it.
Thanks,
what I suggest is that we modify this patch as per Makita-san's suggestion
and proceed with it and the rest of the series. And then come back to the
neigh_tbl_next_index problem.
--
To unsubscribe from this list: send the line "unsubscribe netdev" in
the body of a message to majord...@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
